git.stella-ops.org/docs-archived/product/advisories/14-Jan-2026 - Security gaps in AI-generated code.md

Here’s a quick, practical add-on for your pipelines: simple guardrails for AI‑assisted code so you don’t ship security/IP/license problems without slowing teams down.

---

Why this matters (super short)

AI coding tools boost output but can introduce:

• Security issues (hardcoded secrets, unsafe APIs, prompt‑planted vulns)
• IP leakage (copied proprietary snippets)
• License conflicts (incompatible OSS brought by the model)

---

Three fast checks (under ~1–3s per change)

1. Secrets & Unsafe Patterns

• Detect credentials, tokens, and high‑risk calls (e.g., weak crypto, eval/exec, SQL concat).
• Flag “new secrets introduced” vs “pre‑existing” to reduce noise.
• Require fix or approved suppression with evidence.

2. Attribution & Similarity

• Fuzzy match new/changed hunks against a vetted “allowlist” (your repos) and a “denylist” (company IP you can’t disclose).
• If similarity > threshold to denylist, block; if unknown origin, require justification note.

3. License Hygiene (deps + snippets)

• On dependency diffs, compute SBOM, resolve licenses, evaluate policy matrix (e.g., OK: MIT/BSD/Apache‑2.0; Review: MPL/LGPL; Block: GPL‑3.0 for closed components).
• For pasted code blocks > N lines, enforce “license/attribution comment” presence or ticket link proving provenance.

---

Lightweight policy you can ship today

Policy goals

• Be explainable (every fail has a short reason + link to evidence)
• Be configurable per repo/env
• Support override with audit (who, why, for how long)

Example (YAML)

stellaops:
  ai_code_guard:
    enabled: true
    thresholds:
      similarity_block: 0.92
      similarity_review: 0.80
      max_paste_lines_without_provenance: 12
    licenses:
      allow: [MIT, BSD-2-Clause, BSD-3-Clause, Apache-2.0]
      review: [MPL-2.0, LGPL-2.1, LGPL-3.0]
      block: [GPL-3.0-only, AGPL-3.0-only]
    checks:
      - id: secrets_scan
        required: true
      - id: unsafe_api_scan
        required: true
      - id: snippet_similarity
        required: true
      - id: dep_sbom_license
        required: true
    overrides:
      require_issue_link: true
      max_duration_days: 14

Gate outcomes

• ✅ Pass: merge/release continues
• 🟡 Review: needs approver with role SecurityReviewer
• ⛔ Block: only SecurityOwner can override with issue link + time‑boxed waiver

---

Minimal evidence you store (per change)

• Hashes of changed hunks + similarity scores
• Secret/unsafe findings with line refs
• SBOM delta + license verdicts
• Override metadata (who/why/expiry)

This feeds Stella Ops’ deterministic replay: same inputs → same verdicts → audit‑ready.

---

Drop‑in CI snippets

GitHub Actions

jobs:
  ai-guard:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: stella-ops/cli-action@v1
        with:
          args: guard run --policy .stellaops.yml --format sarif --out guard.sarif
      - name: Upload SARIF
        uses: github/codeql-action/upload-sarif@v3
        with: { sarif_file: guard.sarif }

GitLab CI

ai_guard:
  image: stellaops/cli:latest
  script:
    - stella guard run --policy .stellaops.yml --format gitlab --out guard.json
    - test "$(jq -r .status guard.json)" = "pass"

---

Small UX that wins trust

• Inline PR annotations (secret types, API names, license rule hit)
• One‑click “request waiver” (requires ticket link + expiry)
• Policy badges in PR (“AI Code Guard: Pass / Review / Block”)

---

How this plugs into Stella Ops

• Scanner: run the 3 checks; emit evidence (JSON + DSSE).
• Policy/Lattice Engine: combine verdicts (e.g., Block if secrets OR block‑license; Review if similarity_review).
• Authority: sign the gate result; attach to release attestation.
• Replay: store inputs + rule versions to reproduce decisions exactly.

If you want, I’ll turn this into:

• a ready‑to‑use .stellaops.yml,
• a CLI subcommand spec (stella guard run),
• and UI wireframes for the PR annotations + waiver flow.